osrforge.providers.foundry

The Azure AI Foundry adapter — the only module allowed to import openai/azure.identity.

Minimal but real: the recorded capability probes drive it live and double as its integration test. The probes pinned the two service-behavior choices — the AzureOpenAI client's api-version dialect, and native JSON-schema response format, which the deployment honors — and the validate-and-retry loop keeps the adapter correct even where a service's behavior shifts: the provider owns schema enforcement, and the pipeline never sees invalid data.

FoundryProvider

FoundryProvider(settings: FoundrySettings, client: OpenAI | None = None, sleep: Callable[[float], None] = sleep)

Azure AI Foundry, over the OpenAI-compatible chat surface.

The constructor performs no I/O. Transport policy: bounded exponential backoff on 429/5xx and connection failures, honoring Retry-After; auth and other 4xx failures raise immediately. Images ship as base64 PNG data URLs.

Bind to a deployment.

Parameters:

Name	Type	Description	Default
settings	FoundrySettings	The connection settings.	required
client	OpenAI | None	An injectable OpenAI-compatible client, for tests. When omitted, a real AzureOpenAI client is built from settings (which may raise ProviderError if Entra auth is inferred and azure-identity is missing).	None
sleep	Callable[[float], None]	The backoff sleeper, injectable for tests.	sleep

settings instance-attribute

settings = settings

generate

generate(request: ModelRequest) -> ModelResponse

Run one structured-output completion against the deployment.

Native JSON-schema response format is requested, and the reply is validated against request.schema regardless; an invalid reply is re-prompted with the validation errors appended, at most twice.

Parameters:

Name	Type	Description	Default
request	ModelRequest	The request.	required

Returns:

Type	Description
ModelResponse	The parsed, schema-validated response.

Raises:

Type	Description
ProviderError	On transport, auth, or rate-limit exhaustion.
SchemaValidationError	If no schema-valid reply was obtained within three attempts.

FoundrySettings

Bases: BaseModel

Connection settings for an Azure AI Foundry deployment.

Auth mode is inferred: a key present means key auth; absent means Entra ID via DefaultAzureCredential (which needs the osr-forge[entra] extra).

endpoint instance-attribute

endpoint: str

deployment instance-attribute

deployment: str

api_key class-attribute instance-attribute

api_key: str | None = None

api_version class-attribute instance-attribute

api_version: str = '2024-10-21'

from_env classmethod

from_env() -> FoundrySettings

Build settings from the OSRFORGE_FOUNDRY_* environment variables.

The OSRFORGE_ prefix (not AZURE_OPENAI_*) avoids colliding with other tools' env conventions; the README documents the mapping.

Returns:

Type	Description
FoundrySettings	Settings from OSRFORGE_FOUNDRY_ENDPOINT,
FoundrySettings	OSRFORGE_FOUNDRY_DEPLOYMENT, and (optionally)
FoundrySettings	OSRFORGE_FOUNDRY_API_KEY.

Raises:

Type	Description
ProviderError	If a required variable is missing or empty.